Prepend prefix tokens with the initial timestamp token

faster_whisper/transcribe.py

@@ -684,6 +684,8 @@ def get_prompt(
 
         if prefix:
             prefix_tokens = tokenizer.encode(" " + prefix.strip())
+            if not without_timestamps:
+                prefix_tokens.insert(0, tokenizer.timestamp_begin)
             if len(prefix_tokens) >= self.max_length // 2:
                 prefix_tokens = prefix_tokens[: self.max_length // 2 - 1]
             prompt.extend(prefix_tokens)

requirements.txt

@@ -1,5 +1,5 @@
 av==10.*
-ctranslate2>=3.10,<4
+ctranslate2>=3.17,<4
 huggingface_hub>=0.13
 tokenizers==0.13.*
 onnxruntime>=1.14,<2

tests/test_transcribe.py

@@ -34,6 +34,24 @@ def test_transcribe(jfk_path):
     assert segment.end == segment.words[-1].end
 
 
+def test_prefix_with_timestamps(jfk_path):
+    model = WhisperModel("tiny")
+    segments, _ = model.transcribe(jfk_path, prefix="And so my fellow Americans")
+    segments = list(segments)
+
+    assert len(segments) == 1
+
+    segment = segments[0]
+
+    assert segment.text == (
+        " And so my fellow Americans ask not what your country can do for you "
+        "ask what you can do for your country."
+    )
+
+    assert segment.start == 0
+    assert 10 < segment.end < 11
+
+
 def test_vad(jfk_path):
     model = WhisperModel("tiny")
     segments, info = model.transcribe(